Why should I avoid multiple inheritance in C++?

问题:

Is it a good concept to use multiple inheritance or can I do other things instead?

回答1:

Multiple inheritance (abbreviated as MI) smells, which means that usually, it was done for bad reasons, and it will blow back in the face of the maintainer.

Summary

1. Consider composition of features, instead of inheritance
2. Be wary of the Diamond of Dread
3. Consider inheritance of multiple interfaces instead of objects
4. Sometimes, Multiple Inheritance is the right thing. If it is, then use it.
5. Be prepared to defend your multiple-inherited architecture in code reviews

1. Perhaps composition?

This is true for inheritance, and so, it's even more true for multiple inheritance.

Does your object really needs to inherit from another? A Car does not need to inherit from an Engine to work, nor from a Wheel. A Car has an Engine and four Wheel.

If you use multiple inheritance to resolve these problems instead of composition, then you've done something wrong.

2. The Diamond of Dread

Usually, you have a class A, then B and C both inherit from A. And (don't ask me why) someone then decides that D must inherit both from B and C.

I've encountered this kind of problem twice in 8 eights years, and it is amusing to see because of:

1. How much of a mistake it was from the beginning (In both cases, D should not have inherited from both B and C), because this was bad architecture (in fact, C should not have existed at all...)
2. How much maintainers were paying for that, because in C++, the parent class A was present twice in its grandchild class D, and thus, updating one parent field A::field meant either updating it twice (through B::field and C::field), or having something go silently wrong and crash, later (new a pointer in B::field, and delete C::field...)

Using the keyword virtual in C++ to qualify the inheritance avoids the double layout described above if this is not what you want, but anyway, in my experience, you're probably doing something wrong...

In Object hierarchy, you should try to keep the hierarchy as a Tree (a node has ONE parent), not as a graph.

More about the Diamond (edit 2017-05-03)

The real problem with the Diamond of Dread in C++ (assuming the design is sound - have your code reviewed!), is that you need to make a choice:

• Is it desirable for the class A to exist twice in your layout, and what does it mean? If yes, then by all means inherit from it twice.
• if it should exist only once, then inherit from it virtually.

This choice is inherent to the problem, and in C++, unlike other languages, you can actually do it without dogma forcing your design at language level.

But like all powers, with that power comes responsibility: Have your design reviewed.

3. Interfaces

Multiple inheritance of zero or one concrete classes, and zero or more interfaces is usually Okay, because you won't encounter the Diamond of Dread described above. In fact, this is how things are done in Java.

Usually, what you mean when C inherits from A and B is that users can use C as if it was a A, and/or as if it was a B.

In C++, an interface is an abstract class which has:

1. all its method declared pure virtual (suffixed by = 0) ^{(removed the 2017-05-03)}
2. no member variables

The Multiple inheritance of zero to one real object, and zero or more interfaces is not considered "smelly" (at least, not as much).

More about the C++ Abstract Interface (edit 2017-05-03)

First, the NVI pattern can be used to produce an interface, because the real criteria is to have no state (i.e. no member variables, except this). Your abstract interface's point is to publish a contract ("you can call me this way, and this way"), nothing more, nothing less. The limitation of having only abstract virtual method should be a design choice, not an obligation.

Second, in C++, it makes sense to inherit virtually from abstract interfaces, (even with the additional cost/indirection). If you don't, and the interface inheritance appears multiple time in your hierarchy, then you'll have ambiguities.

Third, object orientation is great, but it is not The Only Truth Out There^TM in C++. Use the right tools, and always remember you have other paradigms in C++ offering different kind of solutions.

4. Do you really need Multiple Inheritance?

Sometimes, yes.

Usually, your C class is inheriting from A and B, and A and B are two unrelated objects (i.e. not in the same hierarchy, nothing in common, different concepts, etc.).

For example, you could have a system of Nodes with X,Y,Z coordinates, able to do a lot of geometric calculations (perhaps a point, part of geometric objects) and each Node is an Automated Agent, able to communicate with other agents.

Perhaps you already have access to two libraries, each with its own namespace (another reason to use namespaces... But you use namespaces, don't you?), one being geo and the other being ai

So you have your own own::Node derive both from ai::Agent and geo::Point.

This is the moment when you should ask yourself if you should not use composition instead. If own::Node is really really both a ai::Agent and a geo::Point, then composition will not do.

Then you'll need multiple inheritance, having your own::Node communicate with other agents according to their position in a 3D space.

(You'll note that ai::Agent and geo::Point are completely, totally, fully UNRELATED... This drastically reduces the danger of multiple inheritance)

Other cases (edit 2017-05-03)

There are other cases:

• using (hopefully private) inheritance as implementation detail
• some C++ idioms like policies could use multiple inheritance (when each part needs to communicate with the others through this)
• the virtual inheritance from std::exception (Is Virtual Inheritance necessary for Exceptions?)
• etc.

Sometimes you can use composition, and sometimes MI is better. The point is: You have a choice. Do it responsibly (and have your code reviewed).

5. So, should I do Multiple Inheritance?

Most of the time, in my experience, no. MI is not the right tool, even if it seems to work, because it can be used by the lazy to pile features together without realizing the consequences (like making a Car both an Engine and a Wheel).

But sometimes, yes. And at that time, nothing will work better than MI.

But because MI is smelly, be prepared to defend your architecture in code reviews (and defending it is a good thing, because if you're not able to defend it, then you should not do it).

回答2:

From an interview with Bjarne Stroustrup:

People quite correctly say that you don't need multiple inheritance, because anything you can do with multiple inheritance you can also do with single inheritance. You just use the delegation trick I mentioned. Furthermore, you don't need any inheritance at all, because anything you do with single inheritance you can also do without inheritance by forwarding through a class. Actually, you don't need any classes either, because you can do it all with pointers and data structures. But why would you want to do that? When is it convenient to use the language facilities? When would you prefer a workaround? I've seen cases where multiple inheritance is useful, and I've even seen cases where quite complicated multiple inheritance is useful. Generally, I prefer to use the facilities offered by the language to doing workarounds

回答3:

There's no reason to avoid it and it can be very useful in situations. You need to be aware of the potential issues though.

The biggest one being the diamond of death:

class GrandParent;
class Parent1 : public GrandParent;
class Parent2 : public GrandParent;
class Child : public Parent1, public Parent2;

You now have two "copies" of GrandParent within Child.

C++ has thought of this though and lets you do virtual inheritence to get around the issues.

class GrandParent;
class Parent1 : public virtual GrandParent;
class Parent2 : public virtual GrandParent;
class Child : public Parent1, public Parent2;

Always review your design, ensure you are not using inheritance to save on data reuse. If you can represent the same thing with composition (and typically you can) this is a far better approach.

回答4:

See w:Multiple Inheritance.

Multiple inheritance has received criticism and as such, is not implemented in many languages. Criticisms includes:

• Increased complexity
• Semantic ambiguity often summarized as the diamond problem.
• Not being able to explicitly inherit multiple times from a single class
• Order of inheritance changing class semantics.

Multiple inheritance in languages with C++/Java style constructors exacerbates the inheritance problem of constructors and constructor chaining, thereby creating maintenance and extensibility problems in these languages. Objects in inheritance relationships with greatly varying construction methods are hard to implement under the constructor chaining paradigm.

Modern way of resolving this to use interface (pure abstract class) like COM and Java interface.

I can do other things in place of this?

Yes, you can. I am going to steal from GoF.

• Program to an Interface, not an Implementation
• Prefer composition over inheritance

回答5:

Public inheritance is an IS-A relationship, and sometimes a class will be an type of several different classes, and sometimes it's important to reflect this.

"Mixins" are also sometimes useful. They are generally small classes, usually not inheriting from anything, providing useful functionality.

As long as the inheritance hierarchy is fairly shallow (as it should almost always be), and well managed, you're unlikely to get the dreaded diamond inheritance. The diamond isn't a problem with all languages that use multiple inheritance, but C++'s treatment of it is frequently awkward and sometimes puzzling.

While I've run into cases where multiple inheritance is very handy, they're actually fairly rare. This is likely because I prefer to use other design methods when I don't really need multiple inheritance. I do prefer to avoid confusing language constructs, and it's easy to construct inheritance cases where you have to read the manual really well to figure out what's going on.

回答6:

You shouldn't "avoid" multiple inheritance but you should be aware of problems that can arise such as the 'diamond problem' ( http://en.wikipedia.org/wiki/Diamond_problem ) and treat the power given to you with care, as you should with all powers.

回答7:

You should use it carefully, there are some cases, like the Diamond Problem, when things can go complicated.

alt text http://www.learncpp.com/images/CppTutorial/Section11/PoweredDevice.gif

回答8:

Every programming language has a slightly different treatment of object-oriented programming with pros and cons. C++'s version places the emphasis squarely on performance and has the accompanying downside that it is disturbingly easy to write invalid code - and this is true of multiple inheritance. As a consequence there is a tendency to steer programmers away from this feature.

Other people have addressed the question of what multiple inheritance isn't good for. But we have seen quite a few comments that more-or-less imply that the reason to avoid it is because it's not safe. Well, yes and no.

As is often true in C++, if you follow a basic guideline you can use it safely without having to "look over your shoulder" constantly. The key idea is that you distinguish a special kind of class definition called a "mix-in"; class is a mix-in if all its member functions are virtual (or pure virtual). Then you are allowed to inherit from a single main class and as many "mix-ins" as you like - but you should inherit mixins with the keyword "virtual". e.g.

class CounterMixin {
    int count;
public:
    CounterMixin() : count( 0 ) {}
    virtual ~CounterMixin() {}
    virtual void increment() { count += 1; }
    virtual int getCount() { return count; }
};

class Foo : public Bar, virtual public CounterMixin { ..... };

My suggestion is that if you intend to use a class as a mix-in class you also adopt a naming convention to make it easy for anyone reading the code to see what's happening & to verify you're playing by the rules of the basic guideline. And you'll find it works much better if your mix-ins have default constructors too, just because of the way virtual base classes work. And remember to make all the destructors virtual too.

Note that my use of the word "mix-in" here isn't the same as the parameterised template class (see this link for a good explanation) but I think it is a fair use of the terminology.

Now I don't want to give the impression that this is the only way to use multiple inheritance safely. It's just one way that is fairly easy to check.

回答9:

At the risk of getting a bit abstract, I find it illuminating to think about inheritance within the frame of category theory.

If we think of all our classes and arrows between them denoting inheritance relations, then something like this

A --> B

means that class B derives from class A. Note that, given

A --> B, B --> C

we say C derives from B which derives from A, so C is also said to derive from A, thus

A --> C

Furthermore, we say that for every class A that trivially A derives from A, thus our inheritance model fulfills the definition of a category. In more traditional language, we have a category Class with objects all classes and morphisms the inheritance relations.

That's a bit of setup, but with that let's take a look at our Diamond of Doom:

C --> D
^     ^
|     |
A --> B

It's a shady looking diagram, but it'll do. So D inherits from all of A, B, and C. Furthermore, and getting closer to addressing OP's question, D also inherits from any superclass of A. We can draw a diagram

C --> D --> R
^     ^
|     |
A --> B
^ 
|
Q

Now, problems associated with the Diamond of Death here are when C and B share some property/method names and things get ambiguous; however, if we move any shared behavior into A then the ambiguity disappears.

Put in categorical terms, we want A, B and C to be such that if B and C inherit from Q then A can be rewritten as as subclass of Q. This makes A something called a pushout.

There is also a symmetric construction on D called a pullback. This is essentially the most general useful class you can construct which inherits from both B and C. That is, if you have any other class R multiply inheriting from B and C, then D is a class where R can be rewritten as as subclass of D.

Making sure your tips of the diamond are pullbacks and pushouts gives us a nice way to generically handle name-clashing or maintenance issues which might arise otherwise.

Note Paercebal's answer inspired this as his admonitions are implied by the above model given that we work in the full category Class of all possible classes.

I wanted to generalize his argument to something which shows how complicated multiple inheritance relationships can be both powerful and non-problematic.

TL;DR Think of the inheritance relationships in your program as forming a category. Then you can avoid Diamond of Doom problems by making multiply-inherited classes pushouts and symmetrically, making a common parent class which is a pullback.

回答10:

We use Eiffel. We have excellent MI. No worries. No issues. Easily managed. There are times to NOT use MI. However, it useful more than people realize because they are: A) in a dangerous language that does not manage it well -OR- B) satisfied with how they've worked around MI for years and years -OR- C) other reasons (too numerous to list I am quite sure--see answers above).

For us, using Eiffel, MI is as natural as anything else and another fine tool in the toolbox. Frankly, we're quite unconcerned that no one else is using Eiffel. No worries. We are happy with what we have and invite you to have a look.

While you're looking: Take special note of Void-safety and the eradication of Null pointer dereferencing. While we're all dancing around MI, your pointers are getting lost! :-)

回答11:

Uses and Abuses of Inheritance.

The article does a great job of explaining inheritance, and it's dangers.

回答12:

Beyond the diamond pattern, multiple inheritance tends to make the object model harder to understand, which in turn increases maintenance costs.

Composition is intrinsically easy to understand, comprehend, and explain. It can get tedious to write code for, but a good IDE (it's been a few years since I've worked with Visual Studio, but certainly the Java IDEs all have great composition shortcut automating tools) should get you over that hurdle.

Also, in terms of maintenance, the "diamond problem" comes up in non-literal inheritance instances as well. For instance, if you have A and B and your class C extends them both, and A has a 'makeJuice' method which makes orange juice and you extend that to make orange juice with a twist of lime: what happens when the designer for 'B' adds a 'makeJuice' method which generates and electrical current? 'A' and 'B' may be compatible "parents" right now, but that doesn't mean they will always be so!

Overall, the maxim of tending to avoid inheritance, and especially multiple inheritance, is sound. As all maxims, there are exceptions, but you need to make sure that there is a flashing green neon sign pointing at any exceptions you code (and train your brain so that any time you see such inheritance trees you draw in your own flashing green neon sign), and that you check to make sure it all makes sense every once in a while.

回答13:

The key issue with MI of concrete objects is that rarely do you have an object that legitimately should "Be an A AND be a B", so it is rarely the correct solution on logical grounds. Far more often, you have an object C that obeys "C can act as an A or a B", which you can achieve via interface inheritance & composition. But make no mistake- inheritance of multiple interfaces is still MI, just a subset of it.

For C++ in particular, the key weakness of the feature isn't the actual EXISTENCE of Multiple Inheritance, but some constructs it allows that are almost always malformed. For example, inheriting multiple copies of the same object like:

class B : public A, public A {};

is malformed BY DEFINITION. Translated into English this is "B is an A and an A". So, even in human language there's a severe ambiguity. Did you mean "B has 2 As" or just "B is an A"?. Allowing such pathological code, and worse making it a usage example, did C++ no favors when it came to making a case for keeping the feature in successor languages.

回答14:

You can use composition in preference to inheritance.

The general feeling is that composition is better, and it's very well discussed.

回答15:

it takes 4/8 bytes per class involved. (One this pointer per class).

This might never be a concern, but if one day you have a micro data structure which is instanced billions of time it will be.